Google’s Gemini 3.5 now includes a new Transcribe feature that converts speech to text in 85 languages while automatically correcting verbal stumbles. The capability aims to clean up filler words, repetitions, and hesitations in real time.
The tool is available within the Gemini 3.5 model, which powers Google’s AI assistant and developer APIs. Users can speak naturally, and the system outputs polished, readable text.
### 85-Language Support
The Transcribe feature covers a broad set of languages, including major world languages and regional dialects. Google says the model was trained on multilingual speech data to handle accents, code-switching, and varying speaking speeds.
- 85 languages are supported, from English and Mandarin to Hindi and Swahili.
- Real-time processing keeps latency low, making it suitable for live transcription.
- Automatic language detection allows seamless switching without manual input.
Each language model was optimized separately, meaning accuracy remains high even for less common languages. Google claims the feature works reliably across noisy environments and informal speech patterns.
### Auto-Correction of Verbal Stumbles
A standout element is the system’s ability to clean up speech disfluencies. Standard speech-to-text tools transcribe every word, including “um,” “uh,” stutters, and false starts. Gemini 3.5 Transcribe filters those out and reconstructs the intended sentence.
- Filler word removal strips “like,” “you know,” and “actually” from the output.
- Repeated phrase correction merges repetitions (e.g., “I I I think” becomes “I think”).
- Punctuation insertion adds commas, periods, and question marks based on prosody.
This auto-correction happens during transcription, not as a separate edit step. The model uses contextual understanding to decide what to keep and what to discard.
“We wanted transcription to feel like reading a well-edited transcript, not a raw recording,” Google’s product lead said. “The AI identifies natural pauses and intention, then writes accordingly.”
The feature does not remove substantive content – only verbal noise. Users can also choose a “verbatim mode” that preserves every utterance if needed.
### Use Cases and Availability
The new transcribe tool is aimed at professionals who need clean meeting notes, journalists transcribing interviews, and anyone who dictates long-form content. It also supports accessibility, helping deaf or hard-of-hearing users get readable captions.
- Google Meet integration allows live captions with auto-correction.
- API access lets developers embed the feature into third-party apps.
- Mobile and desktop versions work across Android, iOS, and web.
The feature is rolling out now in Gemini 3.5 for paid users and developers. Free tier users may get limited access later.
### How It Compares to Previous Models
Earlier versions of Gemini and Google’s speech-to-text API required post-processing to clean transcripts. Gemini 3.5 Transcribe combines recognition and editing in a single pass, reducing latency and complexity.
- Older systems needed separate punctuation and grammar models.
- Gemini 3.5 uses an end-to-end neural architecture that handles both tasks.
- Word error rate is reported significantly lower than prior Google speech models.
The model also handles code-switching – mixing languages in one sentence – better than its predecessors.
### Privacy and Data Handling
Google states that audio input is processed on Google servers, but users in regulated industries can request on-device processing options. For standard cloud use, audio is encrypted and not used for training unless the user opts in.
- Encryption in transit and at rest is standard.
- User-controlled deletion allows erasing transcripts within 30 days.
- Enterprise accounts get additional data residency controls.
Users concerned about privacy can disable voice recording entirely within the Gemini app settings.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.